CN110334764A - Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder - Google Patents
Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder Download PDFInfo
- Publication number
- CN110334764A CN110334764A CN201910600138.8A CN201910600138A CN110334764A CN 110334764 A CN110334764 A CN 110334764A CN 201910600138 A CN201910600138 A CN 201910600138A CN 110334764 A CN110334764 A CN 110334764A
- Authority
- CN
- China
- Prior art keywords
- encoding encoder
- dae
- self
- depth self
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M99/00—Subject matter not provided for in other groups of this subclass
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention proposes a kind of rotary machinery fault diagnosis methods based on deepness auto encoder, it is intended to improve the fault diagnosis precision of rotating machinery, realize step are as follows: acquisition rotating machinery vibrating acceleration time domain signal first obtains training dataset and test data set;Secondly, rolling over cross validation method using K to each activation primitive, pass through the different a series of depth self-encoding encoders of training set training;Then, the depth self-encoding encoder after verifying training is collected by verifying, obtains the precision of each faulty tag;Again, optimal selection parameter is found using trellis search method, depth self-encoding encoder is screened by optimal selection parameter, construct integrated depth self-encoding encoder model;Finally, the prediction label to input sample is obtained, prediction label is mapped back to the fault type of rotating machinery, realizes the fault diagnosis to rotating machinery.
Description
Technical field
The invention belongs to fault diagnosises and signal processing analysis technical field, are related to a kind of rotary machinery fault diagnosis side
Method, and in particular to a kind of rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder, can be used for rolling bearing,
The automatic fault diagnosis of the rotating machineries such as gear-box.
Background technique
Rotating machinery is the most widely used mechanical equipment of industrial circle, is of great significance to socio-economic development.Rotation
Make a connection tool critical component under the bad working environments such as load is big, impact is big, revolving speed is high, ambient noise is big inevitably
Various failures.These failures may cause huge loss and serious casualties.In order to monitor the operation shape of rotating machinery
Condition improves the safety and reliability of rotating machinery, avoids unexpected casualties and economic loss, automatic accurate whirler
Tool fault diagnosis is increasingly valued by people.
Rotary machinery fault diagnosis be mainly by being run to rotating machinery when some dynamic parameters for example temperature, amplitude,
The signals such as displacement are analyzed and processed, and are identified to the data of rotating machinery difference operating condition, to reach the mesh of fault diagnosis
's.In general, the index for evaluating a kind of rotary machinery fault diagnosis method quality has diagnostic accuracy, diagnosis efficiency, robustness, objective
Property etc..
Rotary machinery fault diagnosis method can be divided into conventional fault diagnosis method and intelligent failure diagnosis method.Conventional failure
Diagnostic method is mostly based on signal processing technology using physical model and establishes fault diagnosis model, such as empirical mode decomposition, variation mould
State decomposition, wavelet transformation etc..Original vibration signal often shows complicated, non-linear and more noise in practical engineering applications
Feature needs to rely on advanced signal processing technology for the Accurate Diagnosis of fault type, fault severity level and fault direction,
In addition, the accurate description needs of the rotating machinery performance state under complex working condition extract a large amount of time domains, frequency domain from original signal
And time and frequency domain characteristics, usually from these features screening with diagnosis target correlation by force, more representative feature be one blind
Mesh, subjectivity and time-consuming work, therefore conventional fault diagnosis method carries out feature selecting dependent on expertise, it is objective to lack
Property, it is difficult to the rotating machinery fault state under complex working condition in Practical Project automatically, accurately identify.
Intelligent failure diagnosis method is one kind side got up with sensor and technical development of computer based on data-driven
Method, such as support vector machines, principal component analysis, artificial neural network, stacking-type self-encoding encoder, depth confidence network, convolutional Neural
Network etc..Wherein, although the intelligent diagnosing methods such as support vector machines, principal component analysis and artificial neural network can be got rid of to expert
The dependence of experience realizes the adaptive learning of rotating machinery performance state feature, improves the objectivity of fault diagnosis result,
But these method for diagnosing faults are the intelligent failure diagnosis methods based on shallow-layer feature learning, it is difficult to be mentioned from initial data
Further feature is taken out, causes its feature learning ability weak, fault diagnosis precision is low.
In order to improve the feature learning ability of model, to improve fault diagnosis precision, scholars are proposed with stacking-type
Self-encoding encoder, depth confidence network, convolutional neural networks etc. are the intelligent trouble diagnosis side based on depth characteristic study of representative
Method.However, the shortcomings that deep learning also has oneself, the design comparison of the hyper parameters such as pace of learning, the network structure of model is difficult.
In this case, integrated study is a selection well.Integrated study combines a series of weaker study of learning abilities
Device overcomes the design problem of hyper parameter, achieves preferable learning effect.For example, Shao Haidong et al. in 2018
" the A novel method for delivered on volume 102 of Mechanical Systems and Signal Processing
intelligent fault diagnosis of rolling bearings using ensemble deep auto-
In the article of encoders ", a kind of rolling bearing intelligent failure diagnosis method of integrated depth self-encoding encoder model is proposed, it should
Method acquires the vibration data of rolling bearing first, and divides training set and test set, secondly, being built based on different activation primitives
Vertical integrated depth self-encoding encoder model simultaneously carries out pre-training to model using training set data, utilizes faulty tag on this basis
Network is finely adjusted, finally, being realized using the prediction label of a softmax classifier output test sample according to scene
The bearing vibration time-domain signal acquired in real time diagnoses the malfunction of rolling bearing, is the peace of rotating machinery
Row for the national games and maintenance provide reference.However, this method only considered the overall performance of depth self-encoding encoder, depth self-encoding encoder is not
Integrated Strategy is just directlyed adopt by selection to be combined, and does not consider each depth self-encoding encoder to different faults classification point
The difference of class precision affects the diagnostic accuracy of rolling bearing;Meanwhile each activation primitive only generates a depth from coding
Device, generalization ability is low, and robustness is lower.
Summary of the invention
It is an object of the invention to overcome the problems of the above-mentioned prior art, provide a kind of self-editing based on integrated depth
The rotating machinery intelligent failure diagnosis method of code device, it is intended to improve the fault diagnosis precision of rotating machinery.
Technical thought of the invention is to acquire rotating machinery vibrating acceleration time domain signal first, obtains training dataset
And test data set;Secondly, passing through the different a series of depth self-encoding encoders of training set training to each activation primitive;So
Afterwards, the depth self-encoding encoder after verifying training is collected by verifying, obtains the precision of each faulty tag;Again, it is searched using grid
Suo Fangfa finds optimal selection parameter, is screened by optimal selection parameter to depth self-encoding encoder, constructs integrated depth certainly
Encoder model;Finally, the prediction label to input sample is obtained, prediction label is mapped back to the fault type of rotating machinery,
It realizes to the fault diagnosis of rotating machinery, specifically comprises the following steps:
(1) training dataset X is obtained1With test data set X2:
When (1a) is by the I vibration time-domain signal data randomly selected from rotating machinery database and each vibration
The faulty tag that domain signal data includes is as training dataset X1,The classification of all faulty tags is 1,
2 ..., q ..., Q, wherein Q is the classification sum of faulty tag, I >=2000, and I > > Q, xiIndicate i-th of training sample, yiTable
Show xiFaulty tag;
(1b) makees J vibration time-domain signal data of the rotating machinery to be diagnosed acquired in real time by data collection system
For test data set X2,J≥I/2,xjIndicate j-th of test sample;
(2) multiple depth self-encoding encoders are constructed:
The depth self-encoding encoder that LK output layer is softmax classifier is constructed by L different activation primitives
DAE11,…,DAElk,…,DAELK, first of activation primitive, k-th of depth self-encoding encoder DAE generatedlkComprising N number of from coding
DeviceN-th of self-encoding encoderHidden layer beN-th self-encoding encoderHidden layerNode
Number is h,It is connected with softmax classifier, the DAElkOutput layer number of nodes be o, K indicate first of activation primitive given birth to
At depth self-encoding encoder quantity, l=1,2 ..., L, k=1,2 ..., K, n=1,2 ..., N, L >=2,2≤K < < I, N >=
2, o >=1;
(3) to each depth self-encoding encoder DAElkIt is trained:
The method that (3a) rolls over cross validation using K, by training dataset X1It is K parts the same or similar to be divided into size, often
It is secondary that portion is selected to collect V as verifyingk, remaining K-1 parts is used as training set Tk, select K times altogether, obtain the different training set of K group and test
Card collectionWherein xtIndicate t-th of training sample, ytIndicate xtEvent
Hinder label;
(3b) enables l=1;
(3c) enables k=1;
(3d) enables n=1;
(3e) is by training set TkAs DAElkIn n-th of self-encoding encoderInput, it is rightIt is trained, after being trained
Hidden layer beSelf-encoding encoder
(3f) is by the self-encoding encoder after trainingHidden layerAs DAElkIn (n+1)th self-encoding encoderIt is defeated
Enter, it is rightIt is trained, the hidden layer after being trained isSelf-encoding encoder
Whether (3g) judges n=N true, if so, the depth self-encoding encoder DAE after being trainedlk', otherwise, enable n=n+
1, and execute step (3f);
Whether (3h) judges k=K true, if so, obtaining the depth self-encoding encoder under first of activation primitive after K training
DAEl1',DAEl2',…,DAElK', otherwise, k=k+1 is enabled, and execute step (3d);
Whether (3i) judges l=L true, if so, obtaining the depth self-encoding encoder under L activation primitive after total LK training
DAE11',…,DAElk',…,DAELK', otherwise, l=l+1 is enabled, and execute step (3c);
(4) integrated depth self-encoding encoder model is constructed:
Verifying is collected V by (4a)kAs the depth self-encoding encoder DAE after traininglk' input, calculate DAElk' in q class therefore
Hinder the nicety of grading p of labellk,q, obtain the corresponding nicety of grading matrix P of first of activation primitivel, corresponding point of L activation primitive
Class concentration matrix is P1..., Pl,…,PL, wherein
(4b) searches for P to first of activation primitive, using trellis search methodlIn q class faulty tag correspond to plk,qMost
Excellent screening number ml, Q class faulty tag selects p altogetherlk,qNumber be ml× Q, the corresponding screening number of L activation primitive are
m1,…,ml,…,mL;
(4c) is according to screening number ml, select in the corresponding K depth self-encoding encoder of first of activation primitive to the event of q class
Hinder the highest m of labeling precisionlA depth self-encoding encoder, L activation primitive Q class faulty tag are selectedA depth
Self-encoding encoder;
(4d) is according to screening number mlTo PlIn plk,qIt is screened from big to small, calculates plk,qCorresponding selection parameter
ilk,q, obtain the corresponding selection matrix I of first of activation primitivel, the corresponding selection matrix I of L activation primitive1,…,Il,…,IL,
Wherein
(4e) passes through the depth self-encoding encoder DAE after traininglk', calculate verifying collection VkIn xtBelong to q class faulty tag
Probability value prlk,q, obtain the corresponding probability matrix Pr of first of activation primitivel, the corresponding probability matrix of L activation primitive is
Pr1,…,Prl,…,PrL, wherein
(4f) passes through nicety of grading matrix PlIn plk,q, probability matrix PrlIn prlk,qWith selection matrix IlIn
ilk,q, calculate verifying collection VkIn xtA possibility that belonging to q class faulty tag value PRq, a possibility that Q class faulty tag corresponds to
Value is PR1,…,PRq,…,PRQ, remember PR1,…,PRq,…,PRQMaximum value be PRmax, and by PRmaxCorresponding fault category
Label y 'tAs xtPrediction label;
(4g) is constructedA depth self-encoding encoder, and with xtFor input, with xtPrediction label y 'tIt is defeated
Integrated depth self-encoding encoder model out;
(5) rotary machinery fault diagnosis result is obtained:
(5a) is by test data set X2In xjIntegrated depth self-encoding encoder model is inputted as input vector, calculates xj's
Prediction label y'j, obtain prediction label vector [y1',…,y'j,…,y'J]T;
(5b) is by prediction label y'jWith training dataset X1The fault category for including is mapped, and whirler to be diagnosed is obtained
Malfunction of the tool in different moments.
Compared with the prior art, the invention has the following advantages:
First, the present invention is by integrating depth self-encoding encoder model realization when obtaining rotary machinery fault diagnosis result
, in depth self-encoding encoder integration phase, optimal selection parameter is found using trellis search method, overcome to hyper parameter according to
Classify in the multiple depth self-encoding encoders for relying, and selecting same activation primitive to generate according to selection parameter to all kinds of faulty tags smart
Spend highest preceding several depth self-encoding encoders, it is contemplated that difference of each depth self-encoding encoder to different faults category classification precision
It is different, compared with prior art, effectively increase the fault diagnosis precision of rotating machinery.
Second, the present invention passes through each activation primitive using K folding cross validation in the depth self-encoding encoder training stage
The different multiple depth self-encoding encoders of training set training, improve integrated depth self-encoding encoder model generalization ability, with existing skill
Art is compared, and method for diagnosing faults robustness is improved.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention;
Fig. 2 is the vibration time domain signal waveform schematic diagram of 12 kinds of different faults types of rolling bearing of the embodiment of the present invention;
Fig. 3 is the structural schematic diagram of depth self-encoding encoder of the present invention;
Fig. 4 is the Integrated Strategy schematic diagram for the depth self-encoding encoder that the present invention has identical activation primitive;
Fig. 5 is the over all Integration strategy schematic diagram of all depth self-encoding encoders of the present invention;
Fig. 6 is that the embodiment of the present invention integrates depth self-encoding encoder nicety of grading and selection parameter relation schematic diagram;
Fig. 7 is the rolling bearing intelligent trouble diagnosis result schematic diagram of the embodiment of the present invention.
Specific embodiment
In the following with reference to the drawings and specific embodiments, the present invention is described in further detail:
Referring to Fig.1, the present invention includes the following steps:
Step 1) obtains training dataset X1With test data set X2。
The present invention can be used for the intelligent trouble diagnosis of the rotating machineries such as rolling bearing, gear-box, and the present embodiment is with the axis of rolling
For holding, experimental analysis has been carried out using the bearing fault data of Case Western Reserve University, has been acquired and is rolled by data collection system
Bearing amounts to 12 kinds of fault types, 3600 vibration time-domain signals as data set.It is specific as follows:
The vibration time-domain signal that the present embodiment uses is all from U.S.'s Case Western Reserve University bearing data set.Test bearing master
It to include normal condition, segment sunken four kinds of (BD), outer ring defect (OR) and inner ring defect (IR) fault types.Use electrical spark working
Single Point of Faliure is introduced test bearing by work, and fault diameter includes 0.007,0.014,0.021 and 0.028 inch, totally four kinds of sizes
Type obtains and amounts to 12 kinds of failure classes including different fault types, different faults diameter dimension and different faults orientation
The bearing vibration time-domain signal of type, waveform are as shown in Figure 2.Vibration signal is with the acquisition of 1797rpm motor speed.Each event
Hinder type and acquire 300 samples, randomly select 200 samples and be trained, remaining 100 samples are responsible for what test was proposed
Method.Each sample includes 400 data points, obtains 2400 sample points as training dataset, 1200 sample point conducts
Test data set.
Data description is as shown in table 1.
Rolling Bearing Status in 1 data set of table
Step 2) constructs multiple depth self-encoding encoders:
The depth self-encoding encoder that LK output layer is softmax classifier is constructed by L different activation primitives
DAE11,…,DAElk,…,DAELK, first of activation primitive, k-th of depth self-encoding encoder DAE generatedlkComprising N number of from coding
DeviceN-th of self-encoding encoderHidden layer beN-th self-encoding encoderHidden layerNode
Number is h,It is connected with softmax classifier, the DAElkOutput layer number of nodes be o, K indicate first of activation primitive given birth to
At depth self-encoding encoder quantity, l=1,2 ..., L, k=1,2 ..., K, n=1,2 ..., N, L >=2,2≤K < < I, N >=
2, o >=1;
Referring to Fig. 3, in the present embodiment, depth self-encoding encoder is constructed using 4 kinds of activation primitives, each activation primitive is raw
At 10 depth self-encoding encoders.The equation and derivative of these activation primitives are as shown in table 2.
The equation and derivative of 24 activation primitives of table
It is in order to simplify this method, the architecture of each depth self-encoding encoder is unified to 400-200-100-80.Meanwhile
Depth self-encoding encoder hyper parameter having the same with identical activation primitive is set.The major parameter of depth self-encoding encoder such as table
Shown in 3.
The major parameter of 3 depth self-encoding encoder of table
Step 3) is referring to Fig. 4, to each depth self-encoding encoder DAElkIt is trained:
The method that (3a) uses 10 folding cross validations, by training dataset X1The same or similar 10 parts of size are divided into,
It selects a every time as verifying and collects Vk, remaining 9 parts are used as training set Tk, altogether select 10 times, obtain 10 groups of different training sets and
Verifying collectionWherein xtIndicate t-th of training sample, ytIndicate xt
Faulty tag;
(3b) enables l=1;
(3c) enables k=1;
(3d) enables n=1;
(3e) is by training set TkAs DAElkIn n-th of self-encoding encoderInput, it is rightIt is trained, after being trained
Hidden layer beSelf-encoding encoder
(3f) is by the self-encoding encoder after trainingHidden layerAs DAElkIn (n+1)th self-encoding encoderIt is defeated
Enter, it is rightIt is trained, the hidden layer after being trained isSelf-encoding encoder
Whether (3g) judges n=4 true, if so, the depth self-encoding encoder DAE after being trainedlk', otherwise, enable n=n+
1, and execute step (3f);
Whether (3h) judges k=10 true, if so, obtaining the depth after lower 10 training of first of activation primitive from coding
Device DAEl1',DAEl2',…,DAEl10', otherwise, k=k+1 is enabled, and execute step (3d);
Whether (3i) judges l=4 true, if so, obtaining under 4 activation primitives the depth self-encoding encoder after totally 40 training
DAE11',...,DAElk',…,DAE410', otherwise, l=l+1 is enabled, and execute step (3c);
Step 4) constructs integrated depth self-encoding encoder model:
(4a) collects V referring to Fig. 4, by verifyingkAs the depth self-encoding encoder DAE after traininglk' input, calculate DAElk' in
The nicety of grading p of q class faulty taglk,q, obtain the corresponding nicety of grading matrix P of first of activation primitivel, 4 activation primitives
Corresponding nicety of grading matrix is P1..., Pl,…,P4, wherein
To the depth self-encoding encoder DAE after each traininglk' verified, by verifying collection V accordinglykObtain DAElk'
Nicety of grading vector.Then, with the DAE of identical activation primitivel1',DAEl2',…,DAEl10' precision vector constitute one
Concentration matrix.Finally, the corresponding nicety of grading matrix of 4 activation primitives is P1..., Pl,…,P4Four concentration matrixes.
(4b), to first of activation primitive, searches for P using trellis search method referring to Fig. 5lIn q class faulty tag it is corresponding
plk,qOptimal screening number ml, 12 class faulty tags select p altogetherlk,qNumber be 12ml, the corresponding screening of 4 activation primitives
Number is m1,…,ml,…,m4;
It is screened using nicety of grading of the trellis search method to depth self-encoding encoder, removes lower nicety of grading.
With the corresponding screening number m of sigmoid function1For, choose m1For [1,2,3,4,5,6,7,8,9,10], remaining is optimal sieve
Number is selected, it is as shown in Figure 6 to screen influence of the number value to integrated depth self-encoding encoder nicety of grading, it can be seen that is integrated deep
Self-encoding encoder nicety of grading curve is spent to fluctuate with the variation of screening number.When choosing integrated depth self-encoding encoder nicety of grading highest
MlFor value as optimal screening number, search result is as shown in table 4;
The corresponding screening number of 44 activation primitives of table
(4c) is according to screening number ml, select in the corresponding 10 depth self-encoding encoders of first of activation primitive to the event of q class
Hinder the highest m of labeling precisionlA depth self-encoding encoder, 4 12 class faulty tags of activation primitive are selectedA depth
Self-encoding encoder;
(4d) is according to screening number mlTo PlIn plk,qIt is screened from big to small, calculates plk,qCorresponding selection parameter
ilk,q, obtain the corresponding selection matrix I of first of activation primitivel, the corresponding selection matrix I of 4 activation primitives1,…,Il,…,I4,
Wherein
(4e) passes through the depth self-encoding encoder DAE after traininglk', calculate verifying collection VkIn xtBelong to q class faulty tag
Probability value prlk,q, obtain the corresponding probability matrix Pr of first of activation primitivel, the corresponding probability matrix of 4 activation primitives is
Pr1,…,Prl,…,Pr4, wherein
(4f) passes through nicety of grading matrix PlIn plk,q, probability matrix PrlIn prlk,qWith selection matrix IlIn
ilk,q, calculate verifying collection VkIn xtA possibility that belonging to q class faulty tag value PRq, a possibility that 12 class faulty tags correspond to
Value is PR1,…,PRq,…,PR12, remember PR1,…,PRq,…,PR12Maximum value be PRmax, and by PRmaxCorresponding fault category
Label yt' it is used as xtPrediction label;
(4g) is constructedA depth self-encoding encoder, and with xtFor input, with xtPrediction label yt' it is output
Integrated depth self-encoding encoder model;
(5) rolling bearing fault diagnosis result is obtained:
(5a) is by test data set X2In xjIntegrated depth self-encoding encoder model is inputted as input vector, calculates xj's
Prediction label y'j, obtain prediction label vector [y1',…,y'j,…,y'J]T, as a result as shown in Figure 7;
(5b) is by prediction label y'jWith training dataset X1The fault category for including is mapped, and the axis of rolling to be diagnosed is obtained
Hold the malfunction in different moments.
Below in conjunction with specific experiment, elaborate to technical effect of the invention.
1. experiment condition and content:
Central processing unit be Intel (R) Core (TM) i5-7500 3.40GHZ, memory 16G, WINDOWS7 operation system
On system, rolling bearing intelligent trouble diagnosis result is emulated with MATLAB R2017b software.
2. analysis of experimental results:
Application class diagnostic accuracy Acc of the present invention evaluates and tests the classification diagnosis precision of model, the expression formula of Acc are as follows:
In formula,L(j)For the label predicted j-th of test sample, y(j)Table
Show the physical tags of j-th of test sample.
Performance of the invention, specific comparative experiments are verified using two groups of comparative experimentss are as follows:
First group, the present invention is compared with the intelligent diagnosing method based on single depth self-encoding encoder, comparing result
As shown in table 5.The highest depth self-encoding encoder of accuracy rate is DAE3, accuracy rate 81.58%.All depth self-encoding encoders are put down
Equal precision is 76.89%.The adjustment of any parameter can significantly improve precision, illustrate this method be it is feasible, work is good.
Accuracy obtained by the verifying for the first time of each depth self-encoding encoder of table 5
Second group, the present invention is compared with other intelligent diagnosing methods such as BPNN, SVM, SAE, CNN, comparing result
It is shown in Table 6.
The different Model Diagnosis Comparative results of table 6
According to table 6 as can be seen that compared with other are based on the intelligent failure diagnosis method of single deep learning model, diagnosis
Highest precision is standard CNN, however method of its precision still than proposing in the present invention low 4.5%.
In conclusion the present invention can be improved the fault diagnosis precision of rotating machinery, the dependence to hyper parameter is overcome.
Claims (4)
1. the rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder, it is characterised in that include the following steps:
(1) training dataset X is obtained1With test data set X2:
(1a) believes the I vibration time-domain signal data randomly selected from rotating machinery database and each vibration time domain
The faulty tag that number includes is as training dataset X1,The classification of all faulty tags is 1,2 ...,
Q ..., Q, wherein Q is the classification sum of faulty tag, I >=2000, and I > > Q, xiIndicate i-th of training sample, yiIndicate xi
Faulty tag;
(1b) is using J vibration time-domain signal data of the rotating machinery to be diagnosed acquired in real time by data collection system as survey
Try data set X2,xjIndicate j-th of test sample;
(2) multiple depth self-encoding encoders are constructed:
The depth self-encoding encoder DAE that LK output layer is softmax classifier is constructed by L different activation primitives11,…,
DAElk,…,DAELK, first of activation primitive, k-th of depth self-encoding encoder DAE generatedlkInclude N number of self-encoding encoderN-th of self-encoding encoderHidden layer beN-th self-encoding encoderHidden layerNumber of nodes
For h,It is connected with softmax classifier, the DAElkOutput layer number of nodes be o, K indicate first of activation primitive generated
Depth self-encoding encoder quantity, l=1,2 ..., L, k=1,2 ..., K, n=1,2 ..., N, L >=2,2≤K < < I, N >=2,
o≥1;
(3) to each depth self-encoding encoder DAElkIt is trained:
The method that (3a) rolls over cross validation using K, by training dataset X1It is K parts the same or similar to be divided into size, selects every time
Portion is as verifying collection Vk, remaining K-1 parts is used as training set Tk, select K times altogether, obtain the different training set of K group and verifying collection
{V1,T1},…,{Vk,Tk},…{VK,TK,Wherein xtIndicate t-th of training sample, ytIndicate xtFailure
Label;
(3b) enables l=1;
(3c) enables k=1;
(3d) enables n=1;
(3e) is by training set TkAs DAElkIn n-th of self-encoding encoderInput, it is rightIt is trained, it is hidden after being trained
Hiding layer isSelf-encoding encoder
(3f) is by the self-encoding encoder after trainingHidden layerAs DAElkIn (n+1)th self-encoding encoderInput, it is rightIt is trained, the hidden layer after being trained isSelf-encoding encoder
Whether (3g) judges n=N true, if so, the depth self-encoding encoder DAE after being trainedlk', otherwise, n=n+1 is enabled, and
It executes step (3f);
Whether (3h) judges k=K true, if so, obtaining the depth self-encoding encoder under first of activation primitive after K training
DAEl1',DAEl2',…,DAElK', otherwise, k=k+1 is enabled, and execute step (3d);
Whether (3i) judges l=L true, if so, obtaining the depth self-encoding encoder under L activation primitive after total LK training
DAE11',…,DAElk',…,DAELK', otherwise, l=l+1 is enabled, and execute step (3c);
(4) integrated depth self-encoding encoder model is constructed:
Verifying is collected V by (4a)kAs the depth self-encoding encoder DAE after traininglk' input, calculate DAElk' in q class failure mark
The nicety of grading p of labellk,q, obtain the corresponding nicety of grading matrix P of first of activation primitivel, the corresponding classification essence of L activation primitive
Degree matrix is P1..., Pl,…,PL, wherein
(4b) searches for P to first of activation primitive, using trellis search methodlIn q class faulty tag correspond to plk,qOptimal sieve
Select number ml, Q class faulty tag selects p altogetherlk,qNumber be ml× Q, the corresponding screening number of L activation primitive are m1,…,
ml,…,mL;
(4c) is according to screening number ml, select in the corresponding K depth self-encoding encoder of first of activation primitive to q class faulty tag
The highest m of nicety of gradinglA depth self-encoding encoder, L activation primitive Q class faulty tag are selectedA depth encodes certainly
Device;
(4d) is according to screening number mlTo PlIn plk,qIt is screened from big to small, calculates plk,qCorresponding selection parameter ilk,q,
Obtain the corresponding selection matrix I of first of activation primitivel, the corresponding selection matrix I of L activation primitive1,…,Il,…,IL, wherein
(4e) passes through the depth self-encoding encoder DAE after traininglk', calculate verifying collection VkIn xtBelong to the general of q class faulty tag
Rate value prlk,q, obtain the corresponding probability matrix Pr of first of activation primitivel, the corresponding probability matrix of L activation primitive is
Pr1,…,Prl,…,PrL, wherein
(4f) passes through nicety of grading matrix PlIn plk,q, probability matrix PrlIn prlk,qWith selection matrix IlIn ilk,q, meter
Calculate verifying collection VkIn xtA possibility that belonging to q class faulty tag value PRq, the corresponding likelihood value of Q class faulty tag is
PR1,…,PRq,…,PRQ, remember PR1,…,PRq,…,PRQMaximum value be PRmax, and by PRmaxCorresponding fault category label
y'tAs xtPrediction label;
(4g) is constructedA depth self-encoding encoder, and with xtFor input, with xtPrediction label y'tFor the collection of output
At depth self-encoding encoder model;
(5) rotary machinery fault diagnosis result is obtained:
(5a) is by test data set X2In xjIntegrated depth self-encoding encoder model is inputted as input vector, calculates xjPrediction
Label y'j, obtain prediction label vector [y '1,…,y'j,…,y'J]T;
(5b) is by prediction label y'jWith training dataset X1The fault category for including is mapped, and is obtained rotating machinery to be diagnosed and is existed
The malfunction of different moments.
2. the rotating machinery intelligent failure diagnosis method according to claim 1 based on integrated depth self-encoding encoder, special
Sign is, DAE described in step (4a)lk' in q class faulty tag nicety of grading plk,q, expression formula is
Wherein, y 'tIndicate DAElk' V is collected to verifyingkIn xtPrediction label, ytIt indicates to collect V in verifyingkMiddle xtFaulty tag,
Num () indicates counting function.
3. the rotating machinery intelligent failure diagnosis method according to claim 1 based on integrated depth self-encoding encoder, special
Sign is, calculating r described in step (4c)lk,qCorresponding selection parameter ilk,q, calculation formula are as follows:
Wherein, plk,qIndicate verifying collection VkDepth self-encoding encoder DAE when as input, after traininglk' in q class faulty tag
Nicety of grading.
4. the rotating machinery intelligent failure diagnosis method according to claim 1 based on integrated depth self-encoding encoder, special
Sign is, verifying collection V described in step (4e)kIn xtA possibility that belonging to q class faulty tag value PRq, calculation formula
Are as follows:
Wherein, plk,qIndicate verifying collection VkDepth self-encoding encoder DAE when as input, after traininglk' in q class faulty tag
Nicety of grading, prlk,qIndicate verifying collection VkIn xtBelong to the probability value of q class faulty tag, ilk,qIndicate selection parameter.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910600138.8A CN110334764B (en) | 2019-07-04 | 2019-07-04 | Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910600138.8A CN110334764B (en) | 2019-07-04 | 2019-07-04 | Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110334764A true CN110334764A (en) | 2019-10-15 |
CN110334764B CN110334764B (en) | 2022-03-04 |
Family
ID=68143155
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910600138.8A Active CN110334764B (en) | 2019-07-04 | 2019-07-04 | Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110334764B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110929733A (en) * | 2019-12-09 | 2020-03-27 | 上海眼控科技股份有限公司 | Denoising method and device, computer equipment, storage medium and model training method |
CN111008732A (en) * | 2019-11-21 | 2020-04-14 | 中南大学 | Stack model-based fault prediction method and system |
CN111428788A (en) * | 2020-03-24 | 2020-07-17 | 西安交通大学 | Deep learning-based multi-fault diagnosis method and system for steam turbine generator set rotor |
CN111610026A (en) * | 2020-06-28 | 2020-09-01 | 盐城工学院 | Rotary machine fault diagnosis method based on deep clustering |
CN112183591A (en) * | 2020-09-15 | 2021-01-05 | 上海电机学院 | Transformer fault diagnosis method based on stack sparse noise reduction automatic coding network |
CN112949591A (en) * | 2021-03-31 | 2021-06-11 | 上海电力大学 | Rolling bearing fault diagnosis method based on depth self-encoder |
CN113203566A (en) * | 2021-04-06 | 2021-08-03 | 上海吞山智能科技有限公司 | Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN |
CN113312719A (en) * | 2021-05-27 | 2021-08-27 | 西安电子科技大学 | Rotary machine fault diagnosis method based on class unbalance weight cross entropy |
CN113740667A (en) * | 2021-08-30 | 2021-12-03 | 华北电力大学 | Power grid fault diagnosis method integrating self-encoder and convolutional neural network |
CN114199570A (en) * | 2021-12-13 | 2022-03-18 | 浙江工业大学 | Wind power gearbox bearing fault identification method based on improved variational modal decomposition |
CN114371002A (en) * | 2021-12-30 | 2022-04-19 | 天津理工大学 | Planetary gearbox fault diagnosis method based on DAE-CNN |
CN114707706A (en) * | 2022-03-17 | 2022-07-05 | 广西科学院 | Diesel engine fault diagnosis method based on unbalanced data |
CN114932582A (en) * | 2022-06-16 | 2022-08-23 | 上海交通大学 | Robot small-probability failure prediction method based on Bi-GRU self-encoder |
CN115169505A (en) * | 2022-09-06 | 2022-10-11 | 杭州浅水数字技术有限公司 | Early warning method and early warning system for mechanical fault of special equipment moving part |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080004840A1 (en) * | 2004-04-21 | 2008-01-03 | Pattipatti Krishna R | Intelligent model-based diagnostics for system monitoring, diagnosis and maintenance |
CN108062572A (en) * | 2017-12-28 | 2018-05-22 | 华中科技大学 | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on DdAE deep learning models |
CN108106841A (en) * | 2017-12-21 | 2018-06-01 | 西安交通大学 | Epicyclic gearbox intelligent failure diagnosis method based on built-in encoder signal |
CN109211546A (en) * | 2018-08-28 | 2019-01-15 | 电子科技大学 | Rotary machinery fault diagnosis method based on noise reduction autocoder and incremental learning |
CN109632309A (en) * | 2019-01-17 | 2019-04-16 | 燕山大学 | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning |
-
2019
- 2019-07-04 CN CN201910600138.8A patent/CN110334764B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080004840A1 (en) * | 2004-04-21 | 2008-01-03 | Pattipatti Krishna R | Intelligent model-based diagnostics for system monitoring, diagnosis and maintenance |
CN108106841A (en) * | 2017-12-21 | 2018-06-01 | 西安交通大学 | Epicyclic gearbox intelligent failure diagnosis method based on built-in encoder signal |
CN108062572A (en) * | 2017-12-28 | 2018-05-22 | 华中科技大学 | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on DdAE deep learning models |
CN109211546A (en) * | 2018-08-28 | 2019-01-15 | 电子科技大学 | Rotary machinery fault diagnosis method based on noise reduction autocoder and incremental learning |
CN109632309A (en) * | 2019-01-17 | 2019-04-16 | 燕山大学 | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning |
Non-Patent Citations (2)
Title |
---|
SHAO HAIDONG等: "A novel method for intelligent fault diagnosis of rolling bearings using ensemble deep auto-encoders", 《MECHANICAL SYSTEMS AND SIGNAL PROCESSING》 * |
杨蕊等: "利用 FCKT 以及深度自编码神经网络的滚动轴承故障智能诊断", 《机械工程学报》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111008732B (en) * | 2019-11-21 | 2023-06-20 | 中南大学 | Stack model-based fault prediction method and system |
CN111008732A (en) * | 2019-11-21 | 2020-04-14 | 中南大学 | Stack model-based fault prediction method and system |
CN110929733A (en) * | 2019-12-09 | 2020-03-27 | 上海眼控科技股份有限公司 | Denoising method and device, computer equipment, storage medium and model training method |
CN111428788A (en) * | 2020-03-24 | 2020-07-17 | 西安交通大学 | Deep learning-based multi-fault diagnosis method and system for steam turbine generator set rotor |
CN111610026A (en) * | 2020-06-28 | 2020-09-01 | 盐城工学院 | Rotary machine fault diagnosis method based on deep clustering |
CN112183591A (en) * | 2020-09-15 | 2021-01-05 | 上海电机学院 | Transformer fault diagnosis method based on stack sparse noise reduction automatic coding network |
CN112949591A (en) * | 2021-03-31 | 2021-06-11 | 上海电力大学 | Rolling bearing fault diagnosis method based on depth self-encoder |
CN113203566A (en) * | 2021-04-06 | 2021-08-03 | 上海吞山智能科技有限公司 | Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN |
CN113312719A (en) * | 2021-05-27 | 2021-08-27 | 西安电子科技大学 | Rotary machine fault diagnosis method based on class unbalance weight cross entropy |
CN113740667A (en) * | 2021-08-30 | 2021-12-03 | 华北电力大学 | Power grid fault diagnosis method integrating self-encoder and convolutional neural network |
CN113740667B (en) * | 2021-08-30 | 2022-06-14 | 华北电力大学 | Power grid fault diagnosis method integrating self-encoder and convolutional neural network |
CN114199570A (en) * | 2021-12-13 | 2022-03-18 | 浙江工业大学 | Wind power gearbox bearing fault identification method based on improved variational modal decomposition |
CN114199570B (en) * | 2021-12-13 | 2024-02-02 | 浙江工业大学 | Wind power gear box bearing fault identification method based on improved variation modal decomposition |
CN114371002A (en) * | 2021-12-30 | 2022-04-19 | 天津理工大学 | Planetary gearbox fault diagnosis method based on DAE-CNN |
CN114371002B (en) * | 2021-12-30 | 2024-01-09 | 天津理工大学 | DAE-CNN-based planetary gear box fault diagnosis method |
CN114707706A (en) * | 2022-03-17 | 2022-07-05 | 广西科学院 | Diesel engine fault diagnosis method based on unbalanced data |
CN114932582A (en) * | 2022-06-16 | 2022-08-23 | 上海交通大学 | Robot small-probability failure prediction method based on Bi-GRU self-encoder |
CN114932582B (en) * | 2022-06-16 | 2024-01-23 | 上海交通大学 | Robot small probability failure prediction method based on Bi-GRU self-encoder |
CN115169505A (en) * | 2022-09-06 | 2022-10-11 | 杭州浅水数字技术有限公司 | Early warning method and early warning system for mechanical fault of special equipment moving part |
Also Published As
Publication number | Publication date |
---|---|
CN110334764B (en) | 2022-03-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110334764A (en) | Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder | |
CN106769052B (en) | A kind of mechanical system rolling bearing intelligent failure diagnosis method based on clustering | |
CN110110768B (en) | Rolling bearing fault diagnosis method based on parallel feature learning and multiple classifiers | |
CN109781411B (en) | Bearing fault diagnosis method combining improved sparse filter and KELM | |
CN103728551B (en) | A kind of analog-circuit fault diagnosis method based on cascade integrated classifier | |
CN109655259A (en) | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks | |
CN109186973A (en) | A kind of mechanical failure diagnostic method of unsupervised deep learning network | |
CN109738776A (en) | Fan converter open-circuit fault recognition methods based on LSTM | |
CN106874957A (en) | A kind of Fault Diagnosis of Roller Bearings | |
Xia et al. | Multi-stage fault diagnosis framework for rolling bearing based on OHF Elman AdaBoost-Bagging algorithm | |
CN112765890B (en) | Dynamic domain adaptive network-based multi-working-condition rotating machine residual life prediction method | |
CN108256556A (en) | Wind-driven generator group wheel box method for diagnosing faults based on depth belief network | |
CN206504869U (en) | A kind of rolling bearing fault diagnosis device | |
CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
CN112308147A (en) | Rotating machinery fault diagnosis method based on integrated migration of multi-source domain anchor adapter | |
CN111914883A (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
CN111898644B (en) | Intelligent identification method for health state of aerospace liquid engine under fault-free sample | |
CN113505655A (en) | Bearing fault intelligent diagnosis method for digital twin system | |
CN114298267A (en) | Fault diagnosis method based on bidirectional attention generation countermeasure network and application | |
CN110487547A (en) | Fault Diagnosis of Roller Bearings under variable working condition based on vibrorecord and transfer learning | |
CN113339204A (en) | Wind driven generator fault identification method based on hybrid neural network | |
CN114091504A (en) | Rotary machine small sample fault diagnosis method based on generation countermeasure network | |
CN111783941B (en) | Mechanical equipment diagnosis classification method based on probability confidence convolutional neural network | |
CN117113078A (en) | Small sample bearing fault mode identification method and system based on multi-source data integration | |
CN117475191A (en) | Bearing fault diagnosis method for feature alignment domain antagonistic neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |